From 5cc48232c10fcdc151edf8bb39ad852602d42605 Mon Sep 17 00:00:00 2001 From: kooki3 <1498306179@qq.com> Date: Wed, 17 Dec 2025 17:13:14 +0800 Subject: [PATCH] =?UTF-8?q?=20VisualRobot=5Fwzh=EF=BC=9A=E6=8E=A8=E7=90=86?= =?UTF-8?q?=E8=80=97=E6=97=B6=E8=B0=83=E8=AF=95=E4=BC=98=E5=8C=96?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- VisualRobot/YOLOProcessorORT.cpp | 265 +++++++++---------------------- 1 file changed, 75 insertions(+), 190 deletions(-) diff --git a/VisualRobot/YOLOProcessorORT.cpp b/VisualRobot/YOLOProcessorORT.cpp index e792513..4020c2c 100644 --- a/VisualRobot/YOLOProcessorORT.cpp +++ b/VisualRobot/YOLOProcessorORT.cpp @@ -379,19 +379,22 @@ bool YOLOProcessorORT::DetectObjects(const Mat& frame, vector& cv::Mat rgb; rgb_umat.copyTo(rgb); - // HWC -> CHW + // HWC -> CHW - 使用OpenMP并行化加速 vector inputShape = {1, 3, target_h, target_w}; size_t inputTensorSize = 1 * 3 * target_h * target_w; vector inputTensorValues(inputTensorSize); - int idx = 0; + + // 并行化HWC转CHW的过程 + #pragma omp parallel for collapse(3) for (int c = 0; c < 3; ++c) { for (int y = 0; y < target_h; ++y) { for (int x = 0; x < target_w; ++x) { + size_t idx = c * target_h * target_w + y * target_w + x; Vec3f v = rgb.at(y, x); - inputTensorValues[idx++] = v[c]; + inputTensorValues[idx] = v[c]; } } } @@ -436,22 +439,7 @@ bool YOLOProcessorORT::DetectObjects(const Mat& frame, vector& // 准备调试信息（仅当启用调试时才会使用） YOLODebugInfo debugInfo; - if (enableDebug_) - { - debugInfo.letterbox_r = letterbox_r_; - debugInfo.letterbox_dw = letterbox_dw_; - debugInfo.letterbox_dh = letterbox_dh_; - debugInfo.outputTensorCount = outputTensors.size(); - - // 保存输出形状（在移动张量之前） - for (const auto& out : outputTensors) - { - auto info = out.GetTensorTypeAndShapeInfo(); - debugInfo.outputShapes.push_back(info.GetShape()); - } - } - // debug: print output shapes (useful when user reports mismatch) // convert FIRST output to mat (mimic TestOnnx.py behavior) std::vector singleOuts; // Ort::Value is move-only; avoid copy by moving the element into the new vector @@ -583,99 +571,85 @@ std::vector YOLOProcessorORT::PostProcess(const std::vector cand_class_ids; // 候选框类别ID列表 std::vector> top_candidates; // 候选框信息元组：(索引, 最大sigmoid分数, 原始分数, 类别ID) - // 遍历所有候选框 - 使用OpenMP并行化 - #pragma omp parallel for - for (int c = 0; c < cols; ++c) + // 遍历所有候选框 - 使用OpenMP并行化，优化线程竞争 + #pragma omp parallel { - // 提取中心点坐标和宽高 - float cx = get_val(0, c); - float cy = get_val(1, c); - float w = get_val(2, c); - float h = get_val(3, c); - - // 找到最大的类别分数 - float max_raw = -std::numeric_limits::infinity(); - int best_cls = 0; - for (int k = 0; k < NUM_CLASSES; ++k) + // 每个线程使用局部向量收集结果，减少锁竞争 + vector local_boxes; + vector local_scores; + vector local_class_ids; + + #pragma omp for + for (int c = 0; c < cols; ++c) { - float raw = get_val(4 + k, c); - if (raw > max_raw) + // 提取中心点坐标和宽高 + float cx = get_val(0, c); + float cy = get_val(1, c); + float w = get_val(2, c); + float h = get_val(3, c); + + // 找到最大的类别分数 + float max_raw = -std::numeric_limits::infinity(); + int best_cls = 0; + for (int k = 0; k < NUM_CLASSES; ++k) { - max_raw = raw; - best_cls = k; + float raw = get_val(4 + k, c); + if (raw > max_raw) + { + max_raw = raw; + best_cls = k; + } } - } - - // 按要求将类别分数乘以100用于筛选 - float scaled_conf = max_raw * 100.0f; - - // 应用置信度阈值过滤 - if (scaled_conf < confThreshold_) - { - continue; // 置信度阈值由调用方设置（请注意阈值单位：scaled） - } - // 将 640-scale 像素映射回原图 - double x_orig = (static_cast(cx) - letterbox_dw_) / letterbox_r_; - double y_orig = (static_cast(cy) - letterbox_dh_) / letterbox_r_; - double w_orig = static_cast(w) / letterbox_r_; - double h_orig = static_cast(h) / letterbox_r_; - - // 计算边界框的左上角和右下角坐标 - double x1d = x_orig - w_orig / 2.0; - double y1d = y_orig - h_orig / 2.0; - double x2d = x_orig + w_orig / 2.0; - double y2d = y_orig + h_orig / 2.0; - - // 确保坐标在图像范围内 - int x1 = static_cast(std::max(0.0, std::min(x1d, static_cast(frameSize.width - 1)))); - int y1 = static_cast(std::max(0.0, std::min(y1d, static_cast(frameSize.height - 1)))); - int x2 = static_cast(std::max(0.0, std::min(x2d, static_cast(frameSize.width - 1)))); - int y2 = static_cast(std::max(0.0, std::min(y2d, static_cast(frameSize.height - 1)))); - - // 计算边界框的宽度和高度 - int bw = x2 - x1; - int bh = y2 - y1; - - // 过滤掉无效的边界框 - if (bw <= 0 || bh <= 0) - { - continue; - } + // 按要求将类别分数乘以100用于筛选 + float scaled_conf = max_raw * 100.0f; + + // 应用置信度阈值过滤 + if (scaled_conf < confThreshold_) + { + continue; // 置信度阈值由调用方设置（请注意阈值单位：scaled） + } - // 添加到候选列表 - 使用关键区域确保线程安全 - #pragma omp critical - { - cand_boxes.emplace_back(x1, y1, bw, bh); - cand_scores.push_back(scaled_conf); - cand_class_ids.push_back(best_cls); - } - } - - // 保存候选框信息用于后续分析（顺序不影响结果） - #pragma omp parallel for - for (int c = 0; c < cols; ++c) - { - float cx = get_val(0, c); - float cy = get_val(1, c); - float w = get_val(2, c); - float h = get_val(3, c); - - float max_raw = -std::numeric_limits::infinity(); - int best_cls = 0; - for (int k = 0; k < NUM_CLASSES; ++k) - { - float raw = get_val(4 + k, c); - if (raw > max_raw) + // 将 640-scale 像素映射回原图 + double x_orig = (static_cast(cx) - letterbox_dw_) / letterbox_r_; + double y_orig = (static_cast(cy) - letterbox_dh_) / letterbox_r_; + double w_orig = static_cast(w) / letterbox_r_; + double h_orig = static_cast(h) / letterbox_r_; + + // 计算边界框的左上角和右下角坐标 + double x1d = x_orig - w_orig / 2.0; + double y1d = y_orig - h_orig / 2.0; + double x2d = x_orig + w_orig / 2.0; + double y2d = y_orig + h_orig / 2.0; + + // 确保坐标在图像范围内 + int x1 = static_cast(std::max(0.0, std::min(x1d, static_cast(frameSize.width - 1)))); + int y1 = static_cast(std::max(0.0, std::min(y1d, static_cast(frameSize.height - 1)))); + int x2 = static_cast(std::max(0.0, std::min(x2d, static_cast(frameSize.width - 1)))); + int y2 = static_cast(std::max(0.0, std::min(y2d, static_cast(frameSize.height - 1)))); + + // 计算边界框的宽度和高度 + int bw = x2 - x1; + int bh = y2 - y1; + + // 过滤掉无效的边界框 + if (bw <= 0 || bh <= 0) { - max_raw = raw; - best_cls = k; + continue; } + + // 添加到局部候选列表 + local_boxes.emplace_back(x1, y1, bw, bh); + local_scores.push_back(scaled_conf); + local_class_ids.push_back(best_cls); } + // 合并局部结果到全局列表 #pragma omp critical { - top_candidates.emplace_back(c, 0.0f, max_raw, best_cls); + cand_boxes.insert(cand_boxes.end(), local_boxes.begin(), local_boxes.end()); + cand_scores.insert(cand_scores.end(), local_scores.begin(), local_scores.end()); + cand_class_ids.insert(cand_class_ids.end(), local_class_ids.begin(), local_class_ids.end()); } } @@ -950,7 +924,7 @@ void YOLOProcessorORT::SetDebugOutput(bool enable) * @brief 输出调试信息 * @param debugInfo 调试信息结构体，包含需要输出的各种调试数据 * - * 集中处理所有调试输出，避免调试输出影响核心处理逻辑的性能 + * 集中处理所有调试输出，仅保留必要的计时信息，避免调试输出影响核心处理逻辑的性能 */ void YOLOProcessorORT::DebugOutput(const YOLODebugInfo& debugInfo) { @@ -959,96 +933,7 @@ void YOLOProcessorORT::DebugOutput(const YOLODebugInfo& debugInfo) return; // 调试输出已禁用，直接返回 } - // 输出letterbox信息 - qDebug() << "letterbox r, dw, dh:" << debugInfo.letterbox_r << debugInfo.letterbox_dw << debugInfo.letterbox_dh; - - // 输出模型输出信息 - qDebug() << "ORT returned" << debugInfo.outputTensorCount << "output(s)"; - for (size_t oi = 0; oi < debugInfo.outputTensorCount; ++oi) - { - QString s = "shape:"; - for (auto d : debugInfo.outputShapes[oi]) - { - s += QString::number(d) + ","; - } - qDebug() << " output" << oi << s; - } - - // 输出第一个矩阵的基本信息 - if (!debugInfo.mats.empty()) - { - const Mat &m0 = debugInfo.mats[0]; - qDebug() << "first mat rows,cols:" << m0.rows << m0.cols; - int rshow = std::min(3, m0.rows); - for (int ri = 0; ri < rshow; ++ri) - { - QString rowvals; - for (int ci = 0; ci < std::min(6, m0.cols); ++ci) - { - rowvals += QString::number(m0.at(ri, ci), 'g', 3) + ","; - } - qDebug() << " row" << ri << rowvals; - } - } - - // 输出检测结果的详细信息 - if (!debugInfo.dets.empty()) - { - int R = debugInfo.dets.rows; - int C = debugInfo.dets.cols; - qDebug() << "dets rows,cols:" << R << C; - - // 输出每列的min/max - for (int c = 0; c < C; ++c) - { - double minv, maxv; - cv::minMaxLoc(debugInfo.dets.col(c), &minv, &maxv); - qDebug() << " col" << c << "min" << minv << "max" << maxv; - } - - // 计算每行的最大类别分数和置信度分布 - vector confidences; - confidences.reserve(R); - int cnt_gt_01 = 0, cnt_gt_05 = 0, cnt_gt_001 = 0; - for (int r = 0; r < R; ++r) - { - const float* row = debugInfo.dets.ptr(r); - float obj = (C > 4) ? row[4] : 1.0f; - float maxCls = 0.0f; - for (int cc = 5; cc < C; ++cc) - { - if (row[cc] > maxCls) maxCls = row[cc]; - } - float conf = obj * maxCls; - confidences.push_back(conf); - if (maxCls > 0.1f) - { - cnt_gt_01++; - } - if (maxCls > 0.5f) - { - cnt_gt_05++; - } - if (maxCls > 0.001f) - { - cnt_gt_001++; - } - } - - // 输出top置信度 - vector sorted = confidences; - sort(sorted.begin(), sorted.end(), greater()); - int showN = min((int)sorted.size(), 10); - QString topstr = "top confidences:"; - for (int i = 0; i < showN; ++i) - { - topstr += QString::number(sorted[i], 'g', 6) + ","; - } - qDebug() << topstr; - qDebug() << "counts: >0.001=" << cnt_gt_001 << " >0.1=" << cnt_gt_01 << " >0.5=" << cnt_gt_05; - } - - // 输出处理时间统计 + // 仅输出必要的处理时间统计信息 qDebug() << "YOLO Processing Timing:"; qDebug() << " Preprocess: " << debugInfo.preprocessTime << " ms"; qDebug() << " Inference: " << debugInfo.inferenceTime << " ms"; -- Gitee